Speaker
Description
Modern radio surveys produce large numbers of candidate signals that must be screened to identify new pulsars. This is a rare-event classification problem, since most candidates correspond to noise or radio-frequency interference. In this setting, candidate ranking is as important as binary classification, so predicted probabilities should be reliable for telescope follow-up decisions.
This work studies hybrid quantum-classical logistic regression for calibrated pulsar candidate classification on the imbalanced HTRU-2 dataset. A parameterised quantum circuit is used as a trainable feature map, and Pauli-Z expectation values are passed to a classical logistic layer to estimate the pulsar probability. Three encoding strategies are compared within the same framework: angle encoding, amplitude encoding, and data re-uploading. The quantum logistic models are evaluated against classical logistic regression, SVM-RBF, XGBoost, and a quantum support vector machine reference model.
The evaluation combines discrimination metrics, probability calibration, and runtime. Metrics include PR-AUC, ROC-AUC, Brier score, expected calibration error, reliability diagrams, and Murphy decomposition. Across the tested training-set sizes and circuit depths, angle-encoded quantum logistic regression gives the strongest performance among the quantum logistic regression variants. It remains close to the strongest classical baseline in discrimination and low-false-positive-rate recovery, while giving the lowest calibration error among the compared models.
Shallow angle-encoded quantum logistic regression is overall the best-balanced quantum logistic model for this rare-event screening task. Broadly, the outcomes highlight the role of calibration and operational metrics when quantum machine learning predictions are intended to support scientific decision-making.
| Topic | Hybrid quantum-classical approaches for learning and optimization |
|---|---|
| Additional Topic | Quantum variational circuits and their applications |